do not use backward_val (#3744)

* do not use backward_val

* make variable name more explicit

Author: aloctavodia

pymc3/distributions/transforms.py

@@ -8,7 +8,6 @@
 from .distribution import draw_values
 import numpy as np
 from scipy.special import logit as nplogit
-from scipy.special import expit
 
 
 __all__ = [
@@ -193,9 +192,6 @@ class Log(ElemwiseTransform):
     def backward(self, x):
         return tt.exp(x)
 
-    def backward_val(self, x):
-        return np.exp(x)
-
     def forward(self, x):
         return tt.log(x)
 
@@ -215,9 +211,6 @@ class LogExpM1(ElemwiseTransform):
     def backward(self, x):
         return tt.nnet.softplus(x)
 
-    def backward_val(self, x):
-        return np.log(1 + np.exp(-np.abs(x))) + np.max([x, 0])
-
     def forward(self, x):
         """Inverse operation of softplus.
 
@@ -236,16 +229,12 @@ def jacobian_det(self, x):
 log_exp_m1 = LogExpM1()
 
 
-
 class LogOdds(ElemwiseTransform):
     name = "logodds"
 
     def backward(self, x):
         return invlogit(x, 0.0)
 
-    def backward_val(self, x):
-        return invlogit(x, 0.0)
-
     def forward(self, x):
         return logit(x)
 
@@ -256,7 +245,6 @@ def forward_val(self, x, point=None):
 logodds = LogOdds()
 
 
-
 class Interval(ElemwiseTransform):
     """Transform from real line interval [a,b] to whole real line."""
 
@@ -265,19 +253,12 @@ class Interval(ElemwiseTransform):
     def __init__(self, a, b):
         self.a = tt.as_tensor_variable(a)
         self.b = tt.as_tensor_variable(b)
-        self.a_ = a
-        self.b_ = b
 
     def backward(self, x):
         a, b = self.a, self.b
         r = (b - a) * tt.nnet.sigmoid(x) + a
         return r
 
-    def backward_val(self, x):
-        a, b = self.a_, self.b_
-        r = (b - a) * 1 / (1 + np.exp(-x)) + a
-        return r
-
     def forward(self, x):
         a, b = self.a, self.b
         return tt.log(x - a) - tt.log(b - x)
@@ -304,16 +285,10 @@ class LowerBound(ElemwiseTransform):
 
     def __init__(self, a):
         self.a = tt.as_tensor_variable(a)
-        self.a_ = a
 
     def backward(self, x):
-        a = self.a_
-        r = tt.exp(x) + a
-        return r
-
-    def backward_val(self, x):
         a = self.a
-        r = np.exp(x) + a
+        r = tt.exp(x) + a
         return r
 
     def forward(self, x):
@@ -332,10 +307,10 @@ def jacobian_det(self, x):
 
 
 lowerbound = LowerBound
-'''
+"""
 Alias for ``LowerBound`` (:class: LowerBound) Transform (:class: Transform) class
 for use in the ``transform`` argument of a random variable.
-'''
+"""
 
 
 class UpperBound(ElemwiseTransform):
@@ -345,18 +320,12 @@ class UpperBound(ElemwiseTransform):
 
     def __init__(self, b):
         self.b = tt.as_tensor_variable(b)
-        self.b_ = b
 
     def backward(self, x):
         b = self.b
         r = b - tt.exp(x)
         return r
 
-    def backward_val(self, x):
-        b = self.b_
-        r = b - np.exp(x)
-        return r
-
     def forward(self, x):
         b = self.b
         return tt.log(b - x)
@@ -373,11 +342,10 @@ def jacobian_det(self, x):
 
 
 upperbound = UpperBound
-'''
+"""
 Alias for ``UpperBound`` (:class: UpperBound) Transform (:class: Transform) class
 for use in the ``transform`` argument of a random variable.
-'''
-
+"""
 
 
 class Ordered(Transform):
@@ -389,12 +357,6 @@ def backward(self, y):
         x = tt.inc_subtensor(x[..., 1:], tt.exp(y[..., 1:]))
         return tt.cumsum(x, axis=-1)
 
-    def backward_val(self, y):
-        x = np.zeros(y.shape)
-        x[..., 0] += y[..., 0]
-        x[..., 1:] += np.exp(y[..., 1:])
-        return np.cumsum(x, axis=-1)
-
     def forward(self, x):
         y = tt.zeros(x.shape)
         y = tt.inc_subtensor(y[..., 0], x[..., 0])
@@ -412,10 +374,10 @@ def jacobian_det(self, y):
 
 
 ordered = Ordered()
-'''
+"""
 Instantiation of ``Ordered`` (:class: Ordered) Transform (:class: Transform) class
 for use in the ``transform`` argument of a random variable.
-'''
+"""
 
 
 class SumTo1(Transform):
@@ -430,10 +392,6 @@ def backward(self, y):
         remaining = 1 - tt.sum(y[..., :], axis=-1, keepdims=True)
         return tt.concatenate([y[..., :], remaining], axis=-1)
 
-    def backward_val(self, y):
-        remaining = 1 - np.sum(y[..., :], axis=-1, keepdims=True)
-        return np.concatenate([y[..., :], remaining], axis=-1)
-
     def forward(self, x):
         return x[..., :-1]
 
@@ -500,18 +458,6 @@ def backward(self, y_):
         x = S * yl
         return floatX(x.T)
 
-    def backward_val(self, y_):
-        y = y_.T
-        Km1 = y.shape[0]
-        k = np.arange(Km1)[(slice(None),) + (None,) * (y.ndim - 1)]
-        eq_share = nplogit(1.0 / (Km1 + 1 - k).astype(str(y_.dtype)))
-        z = expit(y + eq_share)
-        yl = np.concatenate([z, np.ones(y[:1].shape)])
-        yu = np.concatenate([np.ones(y[:1].shape), 1 - z])
-        S = np.cumprod(yu, 0)
-        x = S * yl
-        return floatX(x.T)
-
     def jacobian_det(self, y_):
         y = y_.T
         Km1 = y.shape[0]
@@ -526,10 +472,9 @@ def jacobian_det(self, y_):
 stick_breaking = StickBreaking()
 
 
-
 def t_stick_breaking(eps: float) -> StickBreaking:
-    '''Return a new :class:`StickBreaking` transform with specified eps(ilon),
-    instead of the default.'''
+    """Return a new :class:`StickBreaking` transform with specified eps(ilon),
+    instead of the default."""
     return StickBreaking(eps)
 
 
@@ -542,9 +487,6 @@ class Circular(ElemwiseTransform):
     def backward(self, y):
         return tt.arctan2(tt.sin(y), tt.cos(y))
 
-    def backward_val(self, y):
-        return y
-
     def forward(self, x):
         return tt.as_tensor_variable(x)
 
@@ -557,6 +499,7 @@ def jacobian_det(self, x):
 
 circular = Circular()
 
+
 class CholeskyCovPacked(Transform):
     name = "cholesky-cov-packed"
 
@@ -566,10 +509,6 @@ def __init__(self, n):
     def backward(self, x):
         return tt.advanced_set_subtensor1(x, tt.exp(x[self.diag_idxs]), self.diag_idxs)
 
-    def backward_val(self, x):
-        x[..., self.diag_idxs] = np.exp(x[..., self.diag_idxs])
-        return x
-
     def forward(self, y):
         return tt.advanced_set_subtensor1(y, tt.log(y[self.diag_idxs]), self.diag_idxs)
 
@@ -580,6 +519,7 @@ def forward_val(self, y, point=None):
     def jacobian_det(self, y):
         return tt.sum(y[self.diag_idxs])
 
+
 class Chain(Transform):
     def __init__(self, transform_list):
         self.transform_list = transform_list
@@ -603,12 +543,6 @@ def backward(self, y):
             x = transf.backward(x)
         return x
 
-    def backward_val(self, y):
-        x = y
-        for transf in reversed(self.transform_list):
-            x = transf.backward_val(x)
-        return x
-
     def jacobian_det(self, y):
         y = tt.as_tensor_variable(y)
         det_list = []

pymc3/smc/smc.py

@@ -17,7 +17,7 @@
 from ..step_methods.metropolis import MultivariateNormalProposal
 from ..backends.ndarray import NDArray
 from ..backends.base import MultiTrace
-from ..util import get_untransformed_name, is_transformed_name
+from ..util import is_transformed_name
 
 EXPERIMENTAL_WARNING = (
     "Warning: SMC-ABC methods are experimental step methods and not yet"
@@ -128,6 +128,7 @@ def setup_kernel(self):
                 simulator.distribution.function,
                 self.model,
                 self.var_info,
+                self.variables,
                 self.dist_func,
                 self.sum_stat,
             )
@@ -373,7 +374,9 @@ class PseudoLikelihood:
     Pseudo Likelihood
     """
 
-    def __init__(self, epsilon, observations, function, model, var_info, distance, sum_stat):
+    def __init__(
+        self, epsilon, observations, function, model, var_info, variables, distance, sum_stat
+    ):
         """
         epsilon : float
             Standard deviation of the gaussian pseudo likelihood.
@@ -395,9 +398,13 @@ def __init__(self, epsilon, observations, function, model, var_info, distance, s
         self.function = function
         self.model = model
         self.var_info = var_info
+        self.variables = variables
+        self.varnames = [v.name for v in self.variables]
+        self.unobserved_RVs = [v.name for v in self.model.unobserved_RVs]
         self.kernel = self.gauss_kernel
         self.dist_func = distance
         self.sum_stat = sum_stat
+        self.get_unobserved_fn = self.model.fastfn(self.model.unobserved_RVs)
 
         if distance == "absolute_error":
             self.dist_func = self.absolute_error
@@ -409,17 +416,19 @@ def __init__(self, epsilon, observations, function, model, var_info, distance, s
     def posterior_to_function(self, posterior):
         model = self.model
         var_info = self.var_info
-        parameters = {}
+
+        varvalues = []
+        samples = {}
         size = 0
-        for var, values in var_info.items():
-            shape, new_size = values
-            value = posterior[size : size + new_size].reshape(shape)
-            if is_transformed_name(var):
-                var = get_untransformed_name(var)
-                value = model[var].transformation.backward_val(value)
-            parameters[var] = value
+        for var in self.variables:
+            shape, new_size = var_info[var.name]
+            varvalues.append(posterior[size : size + new_size].reshape(shape))
             size += new_size
-        return parameters
+        point = {k: v for k, v in zip(self.varnames, varvalues)}
+        for varname, value in zip(self.unobserved_RVs, self.get_unobserved_fn(point)):
+            if not is_transformed_name(varname):
+                samples[varname] = value
+        return samples
 
     def gauss_kernel(self, value):
         epsilon = self.epsilon